Who decides what your product is made of when an AI answers for you

Five steps, and a claim can fail at any of them

The category talks about this as one property, machine readability, and treats it as though having it were the same as being answered correctly. It is five separate things and each has a different failure.

The practical consequence sets the boundary of this page. Nothing a business publishes controls what an external model or interface says about its products. No structured data, no schema, no feed and no passport confers that, and anybody offering it is selling something that does not exist. What publishing correct, attributable product information does is make the accurate version available and identifiable as yours. That is worth doing on its own terms and it is a narrower promise than the one in circulation.

What the rails actually carry

This is a schema question rather than an opinion, and anybody can check it against the published specification.

The Agentic Commerce Protocol product feed specification, published by OpenAI and Stripe, was read at source on 30 August 2026. What it carries, and what it does not, splits into three states rather than two.

The absences are the load-bearing part and they are consistent across every reading. None of the attributes a textile passport is built from has a field in this specification. Where the specification does address compliance, it addresses disclaimers, warnings and age restrictions, which is the seller's legal notice to a buyer rather than data about what the product is.

A correction to an earlier position, and the reason this page carries the working. Prior research recorded this specification as containing no product identifier at all. That is wrong and it is corrected here. It carries one.

And one thing this page will not state. Two reads of the same specification on the same day returned different accounts of whether a primary material field exists, one describing a capped free-text field and one finding none. That is not a difference this page can resolve from outside, so it publishes neither version. What both reads agree on is the list of absences above, and that is what the argument rests on.

Even on the reading where a material field does exist, it is a short free-text string for a primary material. Fibre composition under textile labelling rules is a quantified percentage per fibre, drawn from a closed list of names. Those are two different objects sharing a word, and where product data stops meaning the same thing takes that failure apart in general. The same attribute name describes different objects in different systems, which is where product data stops meaning the same thing.

The one regulated attribute in production, and why it is there

There is a single place in a major commerce feed where a regulated product attribute genuinely exists and is consumed. A certification attribute, with sub attributes for authority, name, code and value.

Re-verified at source on 28 August 2026, it supports exactly four schemes: an EU energy labelling registry, vehicle carbon dioxide registries in two member states, and a North American energy programme. No textile scheme. No sustainability scheme. No recycled content scheme. No chain of custody scheme.

It exists because energy labelling law compelled the attribute into the feed. Not because a surface asked for it, and not because merchants supplying it were rewarded with visibility.

That is the model, and it is the sentence worth keeping. Law forces the attribute into the feed, and then the surface consumes it. Not the reverse.

The cleanest test of the opposite story failed in public. A major surface announced dozens of new discovery attributes in January 2026. Six generic conversational attributes shipped in May 2026. None of them material, sustainability, certification, provenance or origin.

The argument that does hold

A large commerce platform's agent facing catalogue returns, alongside the merchant's own copy, platform generated fields including product attributes such as material, style and occasion. Its documentation instructs agents to treat those as discovery and merchandising signals rather than as merchant authored source text. The platform sells promoted placement in the same responses.

So: a platform is inferring what a garment is made of and serving that inference to a buying agent, with a disclaimer attached that protects the platform.

Fibre composition is not a soft attribute. It is governed, it is testable against a prescribed method on the finished garment, and a mislabelled garment has been unlawful in the EU for well over a decade. The four rules that already decide what that field may say are set out on fibre composition. The accountability for what is claimed about your product sits with the business that placed it on the market, and a platform's disclaimer about its own generated fields does not move it.

That is an accuracy and liability argument, not an AI visibility argument, and merging the two ruins both.

Two things are not established and they bound how far this can be taken. Whether a merchant opt out from platform generated attributes exists is not documented anywhere we read. How a third party application contributes attributes to that catalogue is also undocumented, and prior work called it the most valuable undocumented question on the platform. Until both are answered, the responsible version of this argument is the one stated above and not a stronger one.

What is rejected, and what would change it

Rejected: that structured or verified sustainability data improves visibility to AI shopping surfaces. Zero primary sources from any AI surface state that verified or regulatory grade attributes are consumed or preferred. One search operator's own documentation states there are no additional requirements to appear in its AI features, no need for new machine readable files and no special structured data. Two independent research streams rejected this proposition and a retest upheld the rejection.

Also rejected: that publishing a machine readable index file improves AI visibility. A study of 137,210 domains found 97 per cent of the existing files received no traffic at all in a month, and no AI bot requests probing for the file on domains that did not have one.

Two named triggers would change this assessment and both are cheap to check.

1. A compliance, certification or provenance field entering a major agentic catalogue schema.
2. A named AI surface documenting that it consumes one.

A weaker third signal would be substantive standards work actually committed at the joint commerce standards workshop, as against being scheduled.

What actually constrains reach today

Crawlability, not semantics. The measured bottleneck is basic.

One vendor panel reports around a third of product pages inaccessible to AI, with the methodology undisclosed, so treat the figure as indicative rather than as a measurement you can rely on. Independent testing found a client rendered page returned 29 characters to a spoofed AI crawler user agent where a server rendered equivalent returned the full content. That is indirect evidence, and neither major model provider makes any public statement about whether they execute JavaScript.

Crawler behaviour differs materially by operator and by bot. Search crawlers honour exclusion rules. User initiated agents largely do not, and no configuration prevents a person asking an assistant to read a public page.

Vendor measurements of citation overlap with organic rankings disagree with each other by wide margins and none is independently audited. They are not quoted here.

The honest advice is the boring advice. Be crawlable, be server rendered, be correct, and hold information that cannot be paraphrased from somewhere else.

The one machine consumption that is real, and it is not shopping

Everything written about machine readability in this category is about shopping agents, which consume no product truth.

The actual machine to machine consumption of a passport identifier is customs. The framework provides for the registry to interconnect with customs systems, and for automated verification at release for free circulation that the registration identifier and the commodity code match what the registry holds. A state system, consuming a product identifier by machine, as a condition of trade.

It binds when a product group is in scope rather than now, and what it does and does not switch on is set out on what happens to a passport at the EU border. It is the clearest case of machine consumption in this whole territory and it is barely discussed, while the discussion that does happen is about surfaces that consume nothing.

What is not being claimed here

It will not claim AI shopping agents are discovering or evaluating passport bearing products. No evidence of any live, publicly resolvable machine readable passport being consumed by any commercial agent was found.

It will not claim merchants are searching in volume for how to make passport data agent ready. Nothing measured supports that.

It will not carry a citation rate target. No reliable measurement method exists that this estate could operate, so a target would be a number with nothing underneath it.

It will not attach a figure to any of the rejections above. An absence of evidence does not acquire a percentage by being written down.

What to do now, in order

Fix crawlability first. It is the only constraint in this territory with a measurement behind it, however imperfect, and it is cheap.

Get identifiers and stable variant identity right. It is the one thing every specification in this space demands, and it is required for reasons that have nothing to do with agents. What allocating an identifier commits you to is a subject in its own right.

Keep a clean merchant feed if you sell through one. It is the prerequisite for agentic checkout on at least one major surface and it is the only place a regulated attribute is consumed anywhere.

Find out what is being said about your products that you did not write. If a platform is generating a material attribute for your garments, that is a claim about your product in a channel you do not control, and you carry the accountability for composition claims regardless of who typed it.

Do not build a sustainability or passport payload for AI surfacing. There are no fields and no evidence. Revisit when one of the two named triggers fires. Composition is the attribute this turns on, and four rules already decide what a declaration may say.

When this page expires

Either named trigger. This page has the shortest half life of anything in this collection and it is written to be replaced rather than defended, and it has already been rewritten once against its own re-check. The feed specification, the four scheme certification list and the Shopify Global Catalog opt-out question should all be re-checked before anybody relies on the position above.

Sources

• Agentic Commerce Protocol product feed specification, OpenAI and Stripe, read at source 30 August 2026
  Specification

  Read for its field set. Carried here as an absence claim for quantified fibre composition, certification, provenance, origin, recycled content and sustainability claims, and as a positive finding that it carries a barcode with the global trade item number among its accepted types. Two reads on the same day disagreed on whether a primary material field exists, so no statement about that field is published. This read corrects a prior research position that the specification carried no product identifier.

  View official source

• A major commerce feed specification's certification attribute, re-verified at source 28 August 2026
  Specification

  Cited for the count and identity of the supported schemes and for its sub attributes.

• A search operator's own developer documentation on AI features
  Vendor documentation

  Cited verbatim in substance for one proposition: that no additional requirements, machine readable files or special structured data are needed to appear in its AI features.

• A large platform's agent facing catalogue documentation
  Platform documentation

  Cited for the existence of platform generated product attributes including material, and for the instruction to agents to treat them as merchandising signals. Whether a merchant opt out exists is not documented in what we read and is recorded as open.

• A study of 137,210 domains on machine readable index file adoption
  Field research

  Cited for two figures with their population stated.

• Ecodesign for Sustainable Products Regulation (EU) 2024/1781, consolidated text
  CELEX 02024R1781-20240628In forceIn force

  Read for the customs interconnection and automated verification provisions only.

  View official source

Sources as at 30 August 2026. Re-check schedule in the body.